← Back to work
LIVELive with real customers

FishTech Sales Agent & Ops

A large language model agent with tool use on FishTech's WhatsApp line, and the operations console behind it. It answers, prices and sends quotations at any hour. Code decides every figure and checks every reply.

Built for
FishTech, Zimbabwe
Role
Design, build and running
Status
Live on WhatsApp

Sample data and an invented conversation. No real customer appears on this page.

Every enquiry needs the same careful work.

FishTech digs and lines fish ponds, sells nets, fingerlings and feed, and delivers across Zimbabwe. Customers write on WhatsApp at any hour, often in Shona or a mix of Shona and English.

Each enquiry needs the same things: a sound answer about fish farming, the right size, a price from the current list, delivery for the distance, and a written quotation. Done by hand, enquiries that arrive at night wait, and quoted prices can drift from the published list.

It sells on WhatsApp. Ops shows everything behind it.

Answers fish-farming questions

From a knowledge sheet the team keeps up to date, in the customer's own language: English, Shona or the mix they write in.

Prices from one list

Pond digging and lining, bird and predator nets, fingerlings, feed and delivery. Sizes are read either way round, and delivery is priced from the road distance to the customer's town.

Sends the quotation

As a WhatsApp message and a PDF, both built by code from the same priced quotation.

Knows when to hand over

Anything it should not answer goes to a person at FishTech, with an alert on WhatsApp and Telegram.

  1. Customer: Mhoro! Ndoda kutanga kurima hove. Pond ye 20 by 10, ndinoisa fingerlings ngani?
  2. Agent: Mhoroi! A 20 m x 10 m lined pond is 200 square metres. Start with about 600 tilapia fingerlings, roughly 3 per square metre, and feed small amounts three times a day while they grow.
  3. Customer: Ok zvakanaka. Ndiitirei quotation, pond ne net ne fingerlings. Ndiri kuChinhoyi.
  4. Agent: Done: pond digging and liner, a bird and predator net, and 600 fingerlings, delivered to Chinhoyi. Here is your quotation.
  5. Agent sends Quotation Q-2026-0142.pdf
  6. Customer: Ndatenda. Ndichakuudzai zuva rekutanga.

The console at ops.fishtech.co.zw holds every chat, client and quotation, from the same code as the WhatsApp commands.

Overview

What needs attention, most urgent first, with the last 7 days and what people ask for.

Inbox

Every chat in full, the reply window for each client, replies from the page and what was agreed beside it.

Clients

Stages, private notes, and exports for a spreadsheet or Google Contacts.

Quotations

Every quotation with its PDF. Approve, reject or withdraw the ones waiting.

Price list and Knowledge

Read straight from the repository and edited there, so the agent and the page never disagree.

Settings

How the agent is running and where its files live.

The model talks. Code decides every figure.

A message arrives at the webhook. One model reads the whole chat and decides what to say, using tools for anything that touches a price or a person. Before a reply leaves, code checks it, and large quotations stop at an approval gate.

A customer's WhatsApp message reaches a FastAPI webhook on Render, which gathers quick messages into one. A language model reads the whole chat and uses tools to prepare a quotation, send the price list or hand the chat to a person. Code then prices every item from the live price list, checks the reply, and holds large quotations for approval. The reply and PDF go out on WhatsApp, alerts go to WhatsApp and Telegram, and chats and quotations are stored on a separate state branch read by the Ops console.

Built from real conversations.

One model, tools for what matters

The first version was a form: once a pond was mentioned, code took over and asked for missing fields, and real customers got looped. Now one model reads the whole chat and decides what to say, and it reaches for a tool for anything that touches a price or a person.

The model never produces a price

Prices reach the customer only as text that code attaches from the price list. A figure in the model's own words blocks the reply.

One price list, shared with the website

The pricing engine reads the same price list that fishtech.co.zw shows, through one feed. A price changed there is live in both places.

Checks before every send

Code blocks invented figures, banned wording, untrue claims about the team and cut-off replies before anything leaves.

Approval for large quotations

Standard quotations go out on their own. Large or non-standard ones wait in Ops until someone approves, rejects or withdraws them.

Quiet when a person steps in

When someone at FishTech takes over a chat with TELL, the agent stays silent there for 24 hours, forwards the customer's messages and sends a reminder after 2 hours.

Every failed real chat is a test

Each real conversation that went wrong became a regression test and a live check against the real model, so the same mistake cannot come back unnoticed.

Fails to a plain, correct reply

If the model fails, or a reply fails the checks twice, the customer gets one fixed, true sentence and the team is alerted. Never half a sentence, never silence.

What it can never do.

  • Conversations, clients and quotations are stored on a separate state branch of a private repository, never in the main code branch.
  • The Ops console sits behind a password, with signed sessions and a limit on failed sign-ins.
  • The agent cannot state anything about the company's legal status. That wording is blocked by code.
  • It never prices a product that is made to order. Those go to a person.
  • Each client shows how long is left in WhatsApp's 24-hour reply window, so the team knows when to call instead.

The stack

Agent and web
  • Python 3.12
  • FastAPI and Uvicorn
  • Large language model agent with tool use
  • A second model as fallback
Pricing and documents
  • Pricing engine in the standard library
  • Price list and knowledge sheet in Markdown
  • fpdf2 for the quotation PDF
  • Road distances from a town list, then OpenStreetMap routing
Channels and running
  • WhatsApp Cloud API
  • Telegram alerts
  • GitHub as the state store
  • Render with an uptime monitor
  • unittest and real-chat evals

Screens

The real Ops page, rendered with invented clients, chats and quotations. Amounts are hidden. Tap a screen to open it.

Want an agent your customers can trust?

Language models where they help, code where the answer must be right. The same pattern fits any business that sells, books or quotes on WhatsApp.

Start a conversation